ClawGym: A Framework for Building Claw-Style AI Agents

ClawGym has been unveiled by researchers as a versatile framework designed for the creation of Claw-style personal agents capable of managing multi-step workflows involving local files, tools, and persistent workspace states. This framework fills a gap in systematic development tools by encompassing the entire agent creation lifecycle, from data synthesis to evaluation and training. Central to this initiative is ClawGym-SynData, a varied dataset comprising 13,500 filtered tasks derived from persona-driven intents and skill-based operations, complemented by realistic mock workspaces and hybrid verification systems. The team developed a series of models known as ClawGym-Agents through supervised fine-tuning on black-box rollout trajectories and investigated reinforcement learning using a streamlined parallel pipeline. This effort seeks to enhance personal agent development with a structured framework for data generation and model training.